Original title: CORAL: An LLM-Native Harness for Production Recommender Systems
Muhammad Rafay Azhar, Yuhang Zhou, Gilbert Jiang, Yuchen Wang +6 · September 2026
Apps use recommender systems to choose what billions of people see online. As user habits and content change, human engineers must constantly run tests to tweak these systems. This testing is slow, so parts of the system get left behind. To fix this, researchers built CORAL, an automated system powered by a large language model, or LLM. Instead of waiting for humans, the LLM agent works in a continuous loop. In each cycle, it checks operating signals, recalls past decisions and their results, and uses tools—like a numerical optimizerA tool that calculates the best adjustments mathematically to meet specific goals.—to tune the recommender while staying within a fixed budget. The system measures the results and uses them to guide the next round. Tested on two large-scale social platforms, CORAL boosted user engagement at no extra cost on one, and cut serving costs without hurting engagement on the other. It kept getting better over time without updating its underlying model parameters.
Imagine tending a garden where plants grow differently each season and you have limited water. Normally, you'd try a watering schedule and adjust manually each week. But if the weather itself changed based on how you watered, you'd need to constantly adapt while staying within your water budget. In this paper, the recommender system is like the garden, and the AI agent is like a gardener who observes growth, remembers past actions, and uses tools to adjust watering within limits, improving the garden over time.
Production recommender systems shape what billions of people see online, but optimizing them is slow and relies on human engineers. This system automates that process, potentially improving user engagement or cutting costs for platforms.
The abstract does not show whether this approach works beyond the two social platforms tested or in different types of recommender systems. It also does not discuss how the LLM itself is maintained or updated over time.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
CORAL: An LLM-Native Harness for Production Recommender Systems — Muhammad Rafay Azhar, Yuhang Zhou, Gilbert Jiang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.